Improving Face Recognition Methods based on POEM Features

Ladislav Lenc

1,2

and Pavel Kr´al

1,2

Dept. of Computer Science & Engineering, Faculty of Applied Sciences, University of West Bohemia,

Plzeˇn, Czech Republic

NTIS - New Technologies for the Information Society, Faculty of Applied Sciences, University of West Bohemia,

Plzeˇn, Czech Republic

Keywords:

POEM, Face Recognition, Local Features, UFI.

Abstract:

POEM descriptors has been successfully used for face recognition. The usual way how the descriptor is

utilized consists in constructing POEM features in the rectangular non-overlapping regions covering the whole

image. The features created in the regions are then concatenated into one long vector representing the face. We

propose an enhancement of this method using automatic key-point identiﬁcation strategies. In our approach,

the image features are created in the detected key-points. We also employ a more complex matching procedure

that compares the features individually. This method is efﬁcient particularly when the number of training

samples is small and therefore neural network based methods fail, because they do not have enough training

data. The proposed approach is evaluated on three standard face corpora. We also study the inﬂuence of

several parameters of the method on the overall performance. The obtained results show that the combination

of POEM features with the automatic point identiﬁcation and a more sophisticated matching algorithm brings

signiﬁcant improvement over the baseline method.

1 INTRODUCTION

Facial recognition is an intensively studied research

ﬁeld that ﬁnds its use in many practical applications.

It is also one of the most useful biometric identiﬁca-

tion methods. There are many successful approaches

based on local features that solve the face recognition

problem.

This work concentrates on the patterns of oriented

edge magnitudes (POEM) descriptors. It has been

proven that POEM has a great ability to capture im-

portant information and it was successfully used for

face recognition (Vu et al., 2012; Lenc, 2016). Sim-

ilarly as other local descriptors, such as local binary

patterns (LBP) and others, image representations are

usually constructed by a concatenation of histograms

of POEM values computed in rectangular image re-

gions. This concept is known as a histogram sequence

(HS). A signiﬁcant improvement of the HS was pro-

posed in (Lenc and Kr´al, 2016)In this approach, LBP

features are created in automatically detected points

and a more sophisticated matching algorithm is used.

The results of this approach show an improvement

over the original methods using LBP and HS.

In this work, we propose a novel face recogni-

tion method which uses automatically detected key-

points together with POEM features. We also aim at a

more detailed evaluation of the inﬂuence of some im-

portant parameters of the key-point detection method.

Compared to the original key-point detection method

we employ also oriented FAST and rotated BRIEF

(ORB) key-point detection algorithm.

We also evaluate several algorithms for key-point

reduction. Some key-point detectors tend to ﬁnd too

many key-points which brings redundancy and in-

creased computational costs. It is thus beneﬁcial to

use algorithms that can reduce the number of key-

points while preserving the informative value of the

key-point set. Another contribution of this work is

the introduction of a new key-point reduction scheme.

The proposed approach is efﬁcient particularly when

the number of training samples is small and therefore

methods based on neural nets fail because of insufﬁ-

cient amount of training data.

The method is evaluated on three standard face

corpora, namely AR, UFI and LFW. AR (Martinez

and Benavente, 1998) database represents a well-

controlled dataset, while UFI (Lenc and Kr´al, 2015)

and LFW (Huang et al., 2007) corpora are much more

challenging because they contain real-world images

with few training examples.

The rest of the paper is organized as follows. The

538

Lenc, L. and Král, P.

Improving Face Recognition Methods based on POEM Features.

DOI: 10.5220/0008950305380545

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 538-545

ISBN: 978-989-758-395-7; ISSN: 2184-433X

following section describes the relevant face recog-

nition methods. Section 3 details the proposed ap-

proach. Section 4 presents the corpora used for evalu-

ation and the following section presents experimental

results realized on these data. The last section con-

cludes the paper and proposes some future research

directions.

2 RELATED WORK

The LBP operator was originally utilized for texture

classiﬁcation in (Ojala et al., 1994). It is a variant

of the texture unit (Wang and He, 1990). Its compu-

tation is based on a small local neighbourhood of a

given pixel. By comparing the central pixel with its

8 neighbouring ones we create an 8 bit code repre-

senting the pixel. The bits in the code are set to 1 if

the values of the given neighbouring pixels are greater

than the value of the central one. The rest of the bits

is set to 0.

The popularity of face recognition methods based

on local descriptors has begun mainly due to the work

of Ahonen et al. (Ahonen et al., 2004; Ahonen et al.,

2006). It has been proven that these simple descrip-

tors, initially used for texture classiﬁcation, are very

useful also in the face recognition ﬁeld. This method

has introduced the histogram sequence (HS) represen-

tation of face images. The utilization of HS ensures

that histograms computed in corresponding parts of

two face images are compared. The reduction of

possible codes using only uniform patterns brought

a speed-up of the method while preserving very good

recognition accuracy.

A plethora of various, more or less sophisticated,

extensions of LBP was proposed in the following

years. We can mention e.g. local ternary patterns

(LTP) (Tan and Triggs, 2010), dynamic threshold lo-

cal binary patterns (DTLBP) (Li et al., 2012) and

completed local binary patterns (CLBP) (Guo et al.,

2010). All of these methods bring some small im-

provements over the basic LBP. The main advantage

is a better handling of lower quality images that are

affected by varying lighting conditions and noise.

Three- and Four-patch LBP variations were pro-

posed in (Wolf et al., 2008). The codes are con-

structed by comparison of three or four patches

respectively. The more sophisticated computation

brings better robustness. The algorithm works very

well on face recognition using the LFW dataset.

Local derivative patterns (LDP) differ in utilizing

features of higher orders than LBP. Again, the main

advantage over LBP is better accuracy in the case of

challenging illumination conditions. A great success

was made by the authors of the POEM descriptor (Vu

et al., 2012). POEM based features outperformed

many other image descriptors and succeeded also in

the face veriﬁcation task.

Further improvements of the basic descriptor

based methods were proposed in (Lenc and Kr´al,

2014) and (Lenc and Kr´al, 2016). The main nov-

elty consists in using automatically detected points

for the feature construction. There is also an im-

provedmatching algorithm that allows reaching better

accuracies in comparison with basic methods using

HS. Another possible improvement lies in weighting.

Each region in the face can have a different weight

which again increases the accuracy. An approach uti-

lizing genetic algorithm to set-up the weights was pro-

posed in (Lenc, 2016).

A thorough description of more algorithms that

were proposed is beyond the scope of this pa-

per, therefore, for the further reading please refer

to (Nanni et al., 2012).

3 POEM-BASED FACE

RECOGNITION

The ﬁrst step in the proposed algorithm is the key-

point identiﬁcation. Two methods for this tasks are

described in Sections 3.1 an 3.2. The second step is

the key-point reduction described in Section 3.3. Fol-

lows the description of the POEM algorithm in Sec-

tion 3.4. The ﬁnal building block of the processing

pipeline is the image matching that is detailed in Sec-

tion 3.5.

3.1 Gabor Wavelet Key-point

Identiﬁcation

We use the Gabor wavelets based method described

in (Lenc and Kr´al, 2014). Gabor ﬁlters utilized in this

work are computed using eqs. (1) and (2) which de-

scribe the real and the imaginary part of the wavelet

respectively (computed in point (r, c)).

g(r, c;λ, θ, ψ, σ, γ) = exp(−

´r+ γ

´c

2σ

)cos(2π

´r

+ ψ)

(1)

g(r, c;λ, θ, ψ, σ, γ) = exp(−

´r+ γ

´c

2σ

)sin(2π

´r

+ ψ)

(2)

where ´r = rcosθ + csinθ, ´c = −rsinθ + ccosθ,

λ is the wavelength of the cosine factor, θ represents

Improving Face Recognition Methods based on POEM Features

539

the orientation of the ﬁlter and ψ is a phase offset,

σ and γ are parameters of the Gaussian envelope,

σ is the standard deviation of the Gaussian and γ

deﬁnes the ellipticity (aspect ratio) of the function.

The magnitude is then computed from the real and

imaginary parts.

We use a set of 40 Gabor ﬁlters that are applied

on the face image. The ﬁlter responses are scanned

using a square sliding window W of the size w × w.

The window centre (r

, c

) is considered to be a key-

point iff:

, c

) = max

(r,c)∈W

(r, c) (3)

, c

) >

wi∗ hi

∑

r=1

∑

c=1

(r, c) (4)

where j = 1, ..., N

is the number of Gabor

ﬁlters) and wi and hi are image width and height

respectively.

We will refer as REAL the method which uses only

the real part and MAGNITUDE the one that computes

the magnitude from the real and the imaginary parts.

3.2 ORB Key-point Identiﬁcation

ORB (Rublee et al., 2011) method was proposed as

an alternative to the patented SIFT and SURF algo-

rithms. It provides key-points that are computed us-

ing the FAST algorithm. BRIEF binary descriptors

are used to create robust features that are successfully

used for image matching and other applications. We

use only the key-pointdetection part of this algorithm.

The FAST algorithm (Rosten and Drummond,

2006) detects corner key-points. Its main strength is

its speed and it is thus often used in real-time appli-

cations. FAST uses a circle with 3 pixel radius and

centre in the candidate key-point. The point P with

the intensity I

is a candidate if the intensities of N

contiguous points on the circle are all:

1. greater then I

+ t

2. smaller then I

− t

where t is a threshold. Parameter N and threshold

t inﬂuence the number of resulting key-point candi-

dates. N is usually set to 12. The ORB algorithm

additionally uses Harris corner ﬁlter to select the best

candidates.

3.3 Key-point Reduction

The key-points detected by one of the described al-

gorithms can be used directly without any reduction.

Unfortunately, the resulting number of key-points and

features is very high. It has been shown (Lenc and

Kr´al, 2014) that this point number can be signiﬁcantly

reduced using clustering with no signiﬁcant informa-

tion loss. Therefore, we evaluate three ways how to

construct the resulting points. All of them are based

on the K-means clustering. We refer as key-points the

points identiﬁed by the particular key-point detection

algorithm. Feature points are the points where fea-

tures are created.

3.3.1 Face Speciﬁc Position

This method constructs the feature points for each im-

age independently. The set of key-points extracted by

the key-point identiﬁcation algorithm is used as the

input of the clustering algorithm. The cluster centres

are then directly used as the feature points. We will

refer this method as FS-POEM.

3.3.2 Person Speciﬁc Position

The second method takes all images of one person

and puts the key-points extracted from all images to-

gether. The points are then clustered and the cluster

centres are used as the feature points. While testing,

we extract key-points from the unknown image and

cluster it to obtain the feature points. This method

will be referred as PS-POEM.

3.3.3 Global Position

The global position method uses the same set of fea-

ture points for each image. A representative subset

of the image gallery is used for key-point extraction.

All resulting points are clustered and the centres are

again used as the feature points. The points are thus

the same for all images. No clustering is performed

for the unknown images. Only the features are cre-

ated in the determined feature points. We will denote

this method as GL-POEM.

3.4 POEM Features

POEM descriptor was proposed in (Vu et al., 2012).

It is based on gradients computed in each image

pixel. The gradients are usually computed by one of

the well-known edge detection convolution operators

such as Sobel or Scharr. The approximation using

these operators allows computing gradients in both x

and y direction and subsequently compute the gradi-

ent magnitude and orientation.

The gradient orientations are then discretized. The

usual number of orientations proposed in the origi-

nal paper is 3. It is denoted d. Vector of the length

d is thus used as a representation of each pixel. It

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

540

is a histogram of gradient values in a small square

neighbourhood of a given pixel called cell. Figure 1

depicts the meaning of cell and block terms.

Figure 1: Computation of POEM descriptor. Squares

around the pixels are called cells while the surroundings

with diameter L is called block. Arrows represent the ac-

cumulated gradients.

The ﬁnal encoding is similar to LBP as depicted in

Figure 2. It is done in a round neighbourhoodwith di-

ameter L called block. The algorithm assigns either 0

or 1 value to the 8 neighbouring pixels by Equation 5.



0 if g

< g

1 if g

≥ g

(5)

where B

is the binary value assigned to the

neighbouring pixel i ∈ {1, .., 8}, g

denotes the

gray-level value of the neighbouring pixel i and g

the gray-level value of the central pixel. The resulting

values are then concatenated into an 8 bit number. Its

decimal representation is used to create the feature

vector.

Figure 2: Computation of the ﬁnal POEM value.

It is computed for each gradient orientation and

thus the descriptor is d times longer than in the case

of LBP.

3.5 Matching Algorithm

In the proposed algorithm we do not concatenate the

feature vectors. Instead, the distance of the feature

sets is computed using the algorithm utilized in (Lenc

and Kr´al, 2016).

The feature vectors are compared using the his-

togram intersection (HI). The advantage of this

method is its simplicity and fast computation. More-

over, in our preliminary experiments it outperformed

some more sophisticated methods such as χ

statistic

in terms of accuracy. HI computation is described by

eq. (6).

HI( f, r) = 1−

∑

min( f

, r

) (6)

where i is the number of histogram bins. This

form is interpreted as a distance measure. 0 value thus

means the same histograms.

The distance of two face representations is com-

puted by eq. (7).

sim(F, R) =

∑

min

∈N( f

)

(HI( f

, r

)) (7)

where N( f

) is the neighbourhood of the feature

deﬁned by the distanceThreshold that speciﬁes

the maximum distance within that the features are

compared. It means that for each feature of the face

F we ﬁnd the closest one within the neighbourhood

N( f

) from the face R. The distance of the two faces

is computed as a sum of these minimum distances.

The recognized face

F is given by the following

equation:

F = argmin

(sim(F, R)) (8)

4 CORPORA

4.1 AR Face Database

AR Face Database

(Martinez and Benavente,

1998) was created at the Univerzitat Auton`oma de

Barcelona. This database contains more than 4,000

colour images of 126 individuals. The images are

stored in a raw format and their size is 768× 576 pix-

els. The individuals are captured under signiﬁcantly

different lighting conditions and with varying expres-

sions. Another characteristic is a possible presence of

glasses or scarf.

4.2 Labeled Faces in the Wild

We use the cropped version of the well-known La-

beled faces in the wild (LFW) dataset (Huang et al.,

http://www2.ece.ohio-state.edu/ aleix/ARdatabase.

html

Improving Face Recognition Methods based on POEM Features

541

2007). This version was ﬁrst utilized in (Sanderson

and Lovell, 2009). The main reason for the cropping

is the presence of a background in the original images

that may add information and improve performance

in some cases. The preprocessing should ensure more

fair conditions for testing.

The extraction method places a bounding box

around the faces and resizes the resulting area to

64 × 64 pixels. The bounding box is placed to the

same location in every image.

We use the identiﬁcation scenario proposed in (Xu

et al., 2014). It uses a subset of 86 people with 11 to

20 images per person. 7 images of each person are

used for the gallery and the rest is used as the probe

set. The total numbers of images are 602 and 649 for

gallery and probe set respectively.

4.3 Unconstrained Facial Images

The Unconstrained facial images (UFI) dataset was

proposed in (Lenc and Kr´al, 2015). It is a real-world

database created from photographs acquired by re-

porters of a news agency. It thus shows signiﬁcant

variances in the image quality, face orientation, face

occlusion etc. The database is designated for the iden-

tiﬁcation task. It comes with two image sets. The

Cropped images dataset contains preprocessed faces

extracted from photographs while the Large images

includes a variable amount of background. We uti-

lize the cropped version while the other partition is

intended to be used with complete face recognition

systems including the face localization stage.

The images have resolution of 128 × 128 pixels.

The total number of individuals is 605. In average

7.1 images of each person are in the gallery set. The

total number of gallery images is 4316. The probe set

contains just one image for every individual.

5 EXPERIMENTS

The ﬁrst experiment was carried out to compare the

key-point identiﬁcation methods. No clustering is

performed in this case and the detected key-points are

directly used as the feature points. The size of the

sliding window is set to 25 according to (Lenc and

Kr´al, 2016). Table 1 shows the comparison of the

methods on all utilized datasets. We report the ac-

curacy and also the number of identiﬁed key-points.

The best results are obtained using the real part of

Gabor wavelets. It is partly due to the larger number

of points that are found by this method. On the other

hand, ORB achieved the worst results. It is evident

mainly on LFW where it also ﬁnds very low number

of key-points. A possible reason for the lower number

of key-poits is lower resolution of the images.

Table 2 compares 3 key-point identiﬁcation meth-

ods with two cluster counts, namely 50 and 100. All

key-point reduction types are examined.

The results show that both Gabor wavelet based

methods have signiﬁcantly better accuracy than ORB.

The real part of Gabor wavelet is mostly sufﬁcient

and achieves slightly better accuracies than magni-

tude. Moreover, it is computationally less expensive.

Based on these experiments, we choose the real part

of Gabor wavelet as the best approach and use it in all

following experiments.

The next experimentperforms a more ﬁne-grained

evaluation of how the number of clusters inﬂuences

the overall accuracy. The three utilized datasets are

tested with all three key-point reduction methods.

100

20 40 60 80 100 120 140

Recognition accuracy

Clusters

Face specific

Person specific

Global

Figure 3: Recognition accuracy in dependence on cluster

count evaluated on AR database.

Figure 3 shows the results for the AR database.

It is obvious that this database is relatively easy and

even a very low number of feature points is enough

to reach high accuracy. Cluster counts higher than

40 are sufﬁcient for all key-point reduction method.

The global position approach allows using even lower

numbers. The highest accuracy of 99.4% is reached

using the Global position. However, the results of the

two other methods are very close and both are higher

than 99%.

Figure 4 reports the results for the challenging

LFW database. In this case, the minimal number of

clusters is around 60. The recognition accuracy then

slightly increases. The highest accuracy of 56.1% is

obtained with person speciﬁc position and 120 clus-

ters.

The curves for the UFI dataset depicted in Fig-

ure 5 have a very similar shape as in the case of the

LFW. The best result is 74.2% for the Global Position

method with 135 clusters.

Table 3 brings the comparison of our results with

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

542

Table 1: Comparison of different key-point determination methods.

Dataset UFI AR LFW

Method ACC (%) Points ACC (%) Points ACC (%) Points

REAL 72.2 330 98.9 413 54.7 352

MAGNITUDE 71.4 275 98.7 333 53.2 290

ORB 68.3 271 95.9 253 41.1 83

Table 2: Comparison of different cluster counts.

Key-point reduction Face Speciﬁc Person Speciﬁc Global

Dataset UFI AR LFW UFI AR LFW UFI AR LFW

Method/clusters ACC (%) ACC (%) ACC (%)

REAL/50 71.1 99.1 52.5 69.1 98.6 51.8 71.7 99.1 51.6

REAL/100 71.7 99.1 54.9 72.4 98.9 55.3 73.4 99.3 53.8

MAGNITUDE/50 71.1 98.7 54.5 65.3 96.3 46.1 71.6 99.1 50.4

MAGNITUDE/100 71.7 98.9 54.4 71.7 98.9 54.2 73.1 99.1 53.9

ORB/50 61.7 94.7 40.8 54.5 92.3 34.1 70.4 98.6 46.8

ORB/100 64.5 95.1 41.0 64.3 95.6 39.0 71.7 98.7 50.5

100

20 40 60 80 100 120 140

Recognition accuracy

Clusters

Face specific

Person specific

Global

Figure 4: Recognition accuracy in dependence on cluster

count evaluated on LFW database.

100

20 40 60 80 100 120 140

Recognition accuracy

Clusters

Face specific

Person specific

Global

Figure 5: Recognition accuracy in dependence on cluster

count evaluated on UFI database.

previously reported LBP based methods. We used the

LBP

8,2

variant. All methods are evaluated with 100

clusters to allow fair comparison.

Table 3: Accuracy (in %) of the proposed methods in com-

parison with LBP-based approaches.

Method / Dataset UFI AR LFW

FS-LBP 64.0 98.9 52.1

PS-LBP 63.3 98.7 51.0

GL-LBP 63.0 98.7 51.8

FS-POEM 71.7 99.1 54.9

PS-POEM 72.4 98.9 55.3

GL-POEM 73.4 99.3 53.8

Table 4: Accuracy of the proposed method on the UFI

dataset in comparison with other approaches reported in the

literature.

Method / Dataset ACC (%)

FS-LBP (Lenc and Kr´al, 2016) 63.96

POEM-HS (Lenc, 2016) 65.95

POEM-HS weighted (Lenc, 2016) 68.93

SIFT (Lenc and Kr´al, 2012) 58.68

M-BNCC (Gaston et al., 2017) 74.55

GL-POEM (proposed) 74.20

The comparison indicates that POEM is superior

mainly on the UFI dataset. The results for the AR

database are comparable while the difference on the

LFW dataset is around 3% in average. We can state

that face speciﬁc method is superior together with

LBP. However, there are differences in the POEM

based method. In this case, person speciﬁc and global

methods perform better than the face speciﬁc one.

Table 4 brings a comparison of our best results ob-

tained on the UFI dataset with the results reported in

the literature. The results for FS-LBP and SIFT are re-

computed because the results reported in the literature

were obtained on older versions of the UFI dataset.

Improving Face Recognition Methods based on POEM Features

543

6 CONCLUSION

In this paper we have proposed an extension of the

POEM-based face recognition method. It combines

automatic detection of feature points and a better

matching algorithm with POEM features. We have

also evaluated several aspects of the method and their

inﬂuence on the resulting accuracy.

The methods were tested on three standard face

corpora. The results are consistently better than those

of the previously published methods using automat-

ically detected points together with LBP features.

Moreover, we were able to reach state-of-the-art ac-

curacy on the UFI dataset.

One of possible improvements is adding weight-

ing also to this method with dynamic feature points.

Based on the results of weighting together with meth-

ods using HS for face representation, it could bring

further increase of the recognition accuracy.

ACKNOWLEDGEMENTS

This work has been partly supported by the project

LO1506 of the Czech Ministry of Education, Youth

and Sports and by the Cross-border Cooperation Pro-

gram Czech Republic - Free State of Bavaria ETS Ob-

jective 2014-2020 (project no. 211).

REFERENCES

Ahonen, T., Hadid, A., and Pietik¨ainen, M. (2004). Face

recognition with local binary patterns. In Computer

vision-eccv 2004, pages 469–481. Springer.

Ahonen, T., Hadid, A., and Pietikainen, M. (2006). Face

description with local binary patterns: Application to

face recognition. IEEE Transactions on Pattern Anal-

ysis & Machine Intelligence, (12):2037–2041.

Gaston, J., Ming, J., and Crookes, D. (2017). Unconstrained

face identiﬁcation with multi-scale block-based corre-

lation. pages 1477–1481.

Guo, Z., Zhang, L., and Zhang, D. (2010). A completed

modeling of local binary pattern operator for texture

classiﬁcation. IEEE Transactions on Image Process-

ing, 19(6):1657–1663.

Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller,

E. (2007). Labeled faces in the wild: A database

for studying face recognition in unconstrained envi-

ronments. Technical report, Technical Report 07-49,

University of Massachusetts, Amherst.

Lenc, L. (2016). Genetic algorithm for weight optimiza-

tion in descriptor based face recognition methods.

In Proceedings of the 8th International Conference

on Agents and Artiﬁcial Intelligence, pages 330–336.

SCITEPRESS-Science and Technology Publications,

Lda.

Lenc, L. and Kr´al, P. (2012). Novel matching methods for

automatic face recognition using sift. In IFIP Inter-

national Conference on Artiﬁcial Intelligence Appli-

cations and Innovations, pages 254–263. Springer.

Lenc, L. and Kr´al, P. (2014). Automatically detected feature

positions for lbp based face recognition. In IFIP In-

ternational Conference on Artiﬁcial Intelligence Ap-

plications and Innovations, pages 246–255. Springer.

Lenc, L. and Kr´al, P. (2015). Unconstrained Facial Im-

ages: Database for face recognition under real-world

conditions. In 14th Mexican International Conference

on Artiﬁcial Intelligence (MICAI 2015), Cuernavaca,

Mexico. Springer.

Lenc, L. and Kr´al, P. (2016). Local binary pattern based

face recognition with automatically detected ﬁdu-

cial points. Integrated Computer-Aided Engineering,

23(2):129–139.

Li, W., Fu, P., and Zhou, L. (2012). Face recognition

method based on dynamic threshold local binary pat-

tern. In Proceedings of the 4th International Confer-

ence on Internet Multimedia Computing and Service,

pages 20–24. ACM.

Martinez, A. and Benavente, R. (1998). The AR Face

Database. Technical report, Univerzitat Auton`oma de

Barcelona.

Nanni, L., Lumini, A., and Brahnam, S. (2012). Survey on

LBP based texturedescriptors for image classiﬁcation.

Expert Systems with Applications, 39(3):3634–3641.

Ojala, T., Pietikainen, M., and Harwood, D. (1994). Per-

formance evaluation of texture measures with classi-

ﬁcation based on kullback discrimination of distribu-

tions. In Pattern Recognition, 1994. Vol. 1-Conference

A: Computer Vision & Image Processing., Proceed-

ings of the 12th IAPR International Conference on,

volume 1, pages 582–585. IEEE.

Rosten, E. and Drummond, T. (2006). Machine learning for

high-speed corner detection. In European conference

on computer vision, pages 430–443. Springer.

Rublee, E., Rabaud, V., Konolige, K., and Bradski, G.

(2011). Orb: An efﬁcient alternative to sift or surf.

In Computer Vision (ICCV), 2011 IEEE international

conference on, pages 2564–2571. IEEE.

Sanderson, C. and Lovell, B. C. (2009). Multi-region proba-

bilistic histograms for robust and scalable identity in-

ference. In International Conference on Biometrics,

pages 199–208. Springer.

Tan, X. and Triggs, B. (2010). Enhanced local texture fea-

ture sets for face recognition under difﬁcult lighting

conditions. IEEE transactions on image processing,

19(6):1635–1650.

Vu, N.-S., Dee, H. M., and Caplier, A. (2012). Face recogni-

tion using the POEM descriptor. Pattern Recognition,

45(7):2478–2488.

Wang, L. and He, D.-C. (1990). Texture classiﬁcation us-

ing texture spectrum. Pattern Recognition, 23(8):905–

910.

Wolf, L., Hassner, T., Taigman, Y., et al. (2008). De-

scriptor based methods in the wild. In Workshop

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

544

on Faces in’Real-Life’Images: Detection, Alignment,

and Recognition.

Xu, Y., Fang, X., Li, X., Yang, J., You, J., Liu, H., and Teng,

S. (2014). Data uncertainty in face recognition. IEEE

transactions on cybernetics, 44(10):1950–1961.

Improving Face Recognition Methods based on POEM Features

545